Gap Scan — 2026-09-08

Scanned

Recent window: 6 New Developments Log/ files (2026-09-05 08:05 through 2026-09-07 22:05); 53 wiki pages with last_updated inside 48 hours. Rotation slice: 12 — sources/ I–P, 172 pages.

Mechanical scans run: bin/lint-scan.py (authoritative broken-link report) and bin/primary-doc-scan.py --days 3 (gap type 9). Eight web operations: two searches and six page fetches, three of them for a single verification.

Gaps actioned (8 of 22 found)

New pages created (live)

• Keith Ellison — Minnesota Attorney General; named on five content pages (X.AI LLC v. Ellison (challenging Minnesota's nudification-technology statute), Memorandum Opinion and Order Denying Preliminary Injunction — X.AI LLC v. Ellison (D. Minn., September 4, 2026), Minnesota HF 1606 (Nudification Technology Prohibition), xAI, State Deepfake Statutes (MN, WA, TX, CA)) and the defendant of record in a case the vault tracks, with no page. Third deferral; created this run because first-party material was finally pulled — the Minnesota Attorney General's own biography page and a multistate letter published by that office. Built from five sources, two of them primary. confidence: medium, honestly: the page rests on documents rather than on any account of Ellison's own AI-policy positions, none of which was located. Records the X.AI v. Ellison and Kohls v. Ellison defendant roles as distinct matters sharing only the defendant and the forum, which the vault had not previously separated. [Cites the queued AG letter inline with a provenance note.]

Pages expanded (live)

• Kids Internet and Digital Safety Act (KIDS Act, H.R. 7757) — the page described H.R. 7757's preemption architecture entirely from post-passage analyst commentary and recorded nothing of the coordinated state opposition to it. Added a ## State attorneys-general opposition section carrying the May 26, 2026 multistate letter's four specific objections with their section citations (§§ 213(c)(2), 220, 243(a)(1), 402(1)(B)), its preemption argument (§§ 106, 221, 237, 244, 304, 407; § 601 and § 601(b)), and its endorsement of S. 1748 by contrast. The AI-specific objection — a "gaping enforcement loophole for AI chat functions that can be construed as 'incidental' to the platform's 'primary purpose'" — is the part the vault most needed and had nowhere. Dating handled carefully: the letter predates the House vote by a month and addresses an earlier draft, so it is presented as prior to, not contradicting, the conflicts-based preemption standard the page attributes to the passed text. 1 new (Source: URL) cite, sources_count 6 → 7, two typed relationships added. Purely additive; nothing removed.
• NIST AI Risk Management Framework (AI RMF 1.0) — highest-in-degree page in the rotation slice (163 inbound) with no ## Relationships section. Added eleven typed links. Also surfaces a placement pattern: Wiki/sources/nist-ai-rmf.md and Wiki/standards/nist-ai-rmf.md are two pages for one instrument, recorded on the new section as related rather than merged.
• New York RAISE Act (S. 8828) — in-degree 140, no ## Relationships. Added eight typed links, including contradicts: [[legislation/eo-14365]], which the page's own prose already carried as the federal-pressure counterweight but had never declared as a relationship.
• Machines of Loving Grace — in-degree 48, no ## Relationships. Added seven typed links. The contradicts: [[concepts/alignment-faking]] edge is qualified in place as a matter of sequence rather than of argument — the essay predates the result and does not engage it.

Link aliases fixed

Zero. bin/lint-scan.py reports 212 distinct broken targets, all referenced exactly once, and 0 referenced 2+ times. Per the settled finding carried since 2026-08-18, the single-reference tail is author surnames, institution shorthand and deliberate forward references, and is not re-audited. The candidate lane the skill directs the run to work — the referenced-2+ list — is empty for the first time in the reports held, which is a change worth naming: it means every broken target with more than one inbound reference has been either created or aliased. The lane produced no candidates this run for that reason, not for want of looking.

Queued — foundational sources

• FCA / Bank of England / HM Treasury, joint statement on frontier AI models and cyber resilience (15 May 2026) — the instrument the 2026-09-07 scan named at score 5 as its first candidate for this run's queue, discovered inside the FCA multi-firm review ingested that day. Verified across two independent primary hosts (fca.org.uk and bankofengland.co.uk), text matching section for section including the footnote; date corroborated four ways in page metadata and again by the Governor's 9 July 2026 letter to the Treasury Committee. Saved: Raw Sources/FCA, Bank of England and Treasury joint statement on frontier AI models and cyber resilience (2026-05-15).md, complete text. Queued: INGEST-fca-boe-hmt-joint-statement-frontier-ai-cyber-2026.md.
• Paglieri, Cross, Genewein, Leibo, Tomasev and Vezhnevets, "A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms" (arXiv:2609.04170) — the only cited primary-document URL across the last six digests with no Raw Sources/ file, and the lead item of the 2026-09-07 evening digest. Verified live at arXiv; identifier, DOI, six-author list and 3 September submission date all confirmed. Saved abstract-level. Queued: INGEST-emergent-cheating-whistleblowing-swarms-2026.md — with a source-fidelity flag, below.
• State Attorneys General, letter opposing the Kids Internet and Digital Safety Act, H.R. 7757 (26 May 2026) — not sought; found while researching Ellison, then confirmed as a gap in its own right when Kids Internet and Digital Safety Act (KIDS Act, H.R. 7757) proved to record no state opposition at all. Verified from a signatory office's own domain. Saved: full letter text. Queued: INGEST-state-ags-letter-opposing-kids-act-2026.md.

Verified and closed without action

• METR Hugging Face incident report — bin/primary-doc-scan.py reported https://metr.org/hugging-face-incident-report-aug-2026.pdf as a possible match under another address. Confirmed: the vault holds it as Raw Sources/Brief Independent Investigation of Agents' Behavior in the OpenAI - Hugging Face Incident (METR and Redwood Research, 2026-08-26).md. Not a gap. This is the second time the scanner's third state has correctly prevented a false-positive capture.

Authenticity-verification failures

None. Three sources pulled, three verified, none rejected.

Two documents were deliberately not pulled on provenance grounds rather than failing verification:

• ESRB/2026/3, Warning of the European Systemic Risk Board of 25 June 2026 on systemic cyber risks stemming from frontier AI models. Named only in a secondary source (a CPD training blog); the canonical host is esrb.europa.eu and the URL was not resolved this run. A warning of an EU body must come from the EU body.
• Governor of the Bank of England, letter to Dame Meg Hillier MP, 9 July 2026. A PDF whose search extraction returned sentences visibly truncated mid-clause — the exact condition under which a hand-transcribed capture introduces drift. Not saved. It is the strongest remaining item in this thread and carries figures the vault does not hold, including Anthropic's claim as of 22 May 2026 to have identified 23,019 candidate software vulnerabilities in open-source software using Claude Mythos Preview.

A contradiction recorded rather than resolved

The multistate AG letter has three different signatory counts across the accounts the vault now holds:

None is asserted anywhere. The discrepancy is written into Keith Ellison, into Kids Internet and Digital Safety Act (KIDS Act, H.R. 7757), into the raw file's frontmatter and into the ingest task. The Minnesota figure is a floor, not a count: the final page's two-column layout collapsed during extraction, so a fortieth or more may be present.

One defect this run introduced and could not repair. The raw file was named Forty State and Territorial Attorneys General - ... before the blocks were counted. This sandbox can create files but not delete or rename them, so the filename asserts a number the capture cannot support. A filename_note in the file's frontmatter and an instruction in the ingest task both say not to quote a count from the filename, and a revisions: entry records the correction. The rename is a curator action.

Deferred backlog (over the daily cap — re-surfaces next run)

• Live-thread page work owned by the nightly fold, not by this skill (score 3 each). The 2026-09-07 22:05 digest is unprocessed at scan time — the seventh consecutive report of an unprocessed evening digest. This is a standing condition of the pipeline. Items left alone: the swarm case study on Unintended coordination between AI agents and Reward Hacking; OpenAI's September 5 statement that its misalignment-incident disclosure practices need to expand and that a framework is coming, for OpenAI; Malaysia's evaluation of Huawei accelerators for a 2 billion ringgit sovereign programme, for US-China AI Competition: Different Races, Different Metrics; the Argentine Patagonia data-centre proposals for Data Center Siting / AI Power Politics; the $229 billion AI-economy revenue estimate for AI Bubble vs. Buildout — Synthesis; and the ARC-AGI-3 harness gap and OpenAI's five altered launch-post metrics for Astra.
• litigation/worker-info-exchange-v-uber (score 4), new this run. A class action for 241,000 Uber drivers over the company's dynamic pay-setting system, filed at Amsterdam's district court and public from 2 September 2026, alleging GDPR breach and earnings suppression. The vault has Automated Decision-Making Technology (ADMT) and Transportation — AI Deployment but no page for the case, no page for Worker Info Exchange, and — notably — no European algorithmic-management litigation coverage at all. Deliberately not created, per the quality gate: the digest carries it entirely through one Guardian report, and a case page built from a single news account of a foreign-court filing would fix that account's framing. The claim document is the right first capture.
• entities/center-for-shared-ai-prosperity (score 4), new this run. Named on Labor Disruption Timelines: Who Predicts What and Why and The 2028 Global Intelligence Crisis, and now the publisher of a 56,000-respondent national survey on AI-disruption policy conducted with Blue Rose Research. Not created: no first-party material describing the organisation was pulled this run. The survey itself is a stronger capture than the org page and should probably be queued first.
• models/grok-imagine (score 3) — third deferral. The product at issue in X.AI LLC v. Ellison and in the CSAM suits, reached across the vault only through Grok (xAI) and litigation pages. Three runs have now deferred it for the same reason: a model page needs capability and release material, and every run's budget has gone elsewhere. It is close to structural. Recommend it be given a direct instruction or handed to model-robustness-check, which is better shaped for it.
• Six further slice-12 sources/ pages with no ## Relationships (score 1–2 each), new this run. Of 172 sources/ I–P pages, 9 lack the section — against 15 of 374 in slice 11. Three were worked. Remaining, by in-degree: Japan AI Promotion Act — Source Summary (28, and confidence: high on 2 sources), Open Problems in Technical AI Governance (17), Premature Antitrust Standards in Algorithmic Pricing (14), OECD — Assessing Potential Future AI Risks, Benefits, and Policy Imperatives (13), Open Problems in Emergent Misalignment (8), Introduction to AI Safety, Ethics and Society (4). At 9 of 172, this slice is the best-wired audited to date on a per-page basis.
• The confidence: high on thin sourcing pattern — sixth consecutive report, now in an eighth folder. Slice 12 returns 102 of 172 pages at confidence: high with sources_count ≤ 1, led by Paris AI Action Summit Declaration (2025) (in-degree 59), MIT NANDA — The GenAI Divide (State of AI in Business 2025) (35), NIST AI Agent Standards Initiative (Feb 2026) (31) and ICRC Position on Autonomous Weapon Systems (source summary) (31) — none reinforced since 2026-06-06, so all three decay windows have lapsed on every one of them. Named as a pattern on 2026-09-03 (legislation/), 09-04 (litigation/, government/), 09-05 (models/), 09-06 (analysis/) and 09-07 (sources/ A–H). This skill's 6–10 daily cap structurally cannot close it at one page a day, and six reports of saying so have not changed that. The recommendation from 2026-09-06 stands and is now overdue: schedule a confidence-audit pass as its own operation.
• Seven slice-12 sources/ pages under 350 words, none at high in-degree: AI Accountability Agenda (Senator Edward J. Markey, July 2026) (225 words), Institutional Sovereignty in the Age of AI (Palantir, July 2026) (295), The Data Center Water Problem Is Soluble (Robin Gaster, ITIF, July 2026) (299), IEA — Energy and AI (329), Protecting Wellbeing Of Users (177), Privacy Without Monopoly: Data Protection and Interoperability — Bennett Cyphers and Cory Doctorow (EFF, 2021) (323), Law-Following AI: Designing AI Agents to Obey Human Laws (O'Keefe, Ramakrishnan, Tay & Winter, 2025) (299). Four carry confidence: high. These are source-robustness-check work, not this skill's.
• Two chain-of-thought monitoring papers, arXiv:2507.11473 and arXiv:2503.11926 (score 4) — second deferral. Cited by Pachocki as the technical basis for OpenAI's primary monitoring bet. Still not held.
• OpenAI June 2026 Threat Report PDF (score 4) — second deferral. Epoch AI AI-R&D task taxonomy (score 3) — second deferral. ARC Prize Astra evaluation post (score 3) — second deferral, and it acquired a second reason to want it this run: the 09-07 morning digest carries the 96.7%-versus-62.7% harness gap and the five altered launch-post metrics, all secondhand.
• Supreme Court of India, Vijay Ghanshyam Gadiya v. Union of India (score 5) — second deferral. Still a scope decision more than a capture task; the vault has no Indian AI litigation coverage to land it in.
• entities/adam-thierer (score 3) — eleventh consecutive deferral, structural.
• Three slice-3 pages, tenth deferral — structural: Standards-Industrial Complex, Three Theories of Victory (US / EU / China AI Governance), Transaction Costs (Coase) and the Agentic Economy. Ten runs without reaching the top ten. They need a direct instruction or a decision that they are not gaps.
• Slice carryover. Slice-4 (eighth): American Civil Liberties Union (ACLU), Access Now. Slice-5 (seventh): Helen Toner, International Monetary Fund (IMF), Henry Farrell. Slice-6 (sixth): Stanford Center for Research on Foundation Models (CRFM), Robert Chesney, Sen. Elizabeth Warren. Slice-7 (sixth): the fifteen high-confidence-thin-sourcing legislation/ pages. Slice-8 (fifth): Getty Images v. Stability AI (US), ISO/IEC 42005 — AI Impact Assessment, Tremblay v. OpenAI, GEMA v. Suno (Munich Regional Court), Sony Music v. Udio (second suit, 2026) — the last two still need a docket check, and the CourtListener MCP is unauthenticated for the sixth consecutive run, now a standing capability gap that will not resolve without the curator authorising the connector. Slice-9 (fourth): Helix (Figure AI Vision-Language-Action model), S1 (Skild AI), Llama 4 (Meta), Llama 3 (Meta). Slice-10 (third): Missing Coverage: Companies, People, Models, Legislation, Concepts (revised) and Crystallize Session: Wiki Gaps & User Knowledge Map (2026-04-13) — April 2026 session artifacts sitting in mainspace, a placement decision the v4.3 content/operational split did not reach. Slice-11 (second): thirteen sources/ A–H pages without ## Relationships.
• Twenty-one source-robustness-check candidates across slices 11–13 — fifteenth consecutive report, structural. Led by California SB 53 — Transparency in Frontier AI Act (in-degree 254) and TAKE IT DOWN Act — Source Summary (97). Out of this skill's lane.
• Date-stamped headers in lint-report.md — twenty-second cycle, and the count moved: 26 → 26 held, with six self-references unchanged. bin/lint-scan.py reports one unprocessed developments-log file.
• Carried unchanged. entities/nightingale; Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts, Thomas Larsen — all four named again in the 09-07 evening digest as authors of the German wiki account, which strengthens the case slightly; the CCW autonomous-weapons consensus text of 5 September 2026; legislation/caadca; entities/david-feith and entities/hudson-institute; Marvell; California AB 1856; the Texas OAG / Meta Webster–Mahoney exchange; the draft executive order on a frontier-developer self-regulatory organization; the AI Evaluator Forum / AEF-1 standard; ExploitBench flag-level disclosure; the "Carolina Principles"; OpenAI's "Our agreement with the Department of War"; the House E&C markup memo; the New York City Public Schools generative-AI policy; the Agentic Commerce Protocol / Model Context Protocol placement question; OpenAI's August 31 response to Rep. Casar (Google Drive host, curator decision); the July 2026 statement from more than 1,000 AI-firm employees; the Swiss Re data-center insurance report (no URL to attempt). Newly joined: CMORG "Firm Guidance For Frontier AI" (June 2026), NCSC frontier-AI guidance and the Minnesota AG Emerging Technology report (2025), all named in documents captured this run and none held.

Run integrity

• No broken outbound links introduced. All five touched content pages and all six queue files resolved against the full 1,800+ page set after writing: zero unresolved on every content page. Three unresolved targets were caught in the queue files and fixed — concepts/ai-cyber-capabilities → AI and Cybersecurity and Autonomous cyber-agents; concepts/multi-agent-systems → Unintended coordination between AI agents, which is the better target anyway; and entities/rob-bonta, which does not exist, rewritten as prose naming him as a page candidate.
• Style scan clean. All five touched content pages checked programmatically against the banned-header, hyperbole, self-reference, date-led-header and standing-Predictions patterns: zero hits on each.
• Revision fidelity. Every content edit was purely additive except three frontmatter values (last_updated on four pages, sources_count on one). No prose, no citation and no wikilink was removed from any page. The ## Relationships sections added to the three sources/ pages introduce no claim not already carried in the page's own body.
• Confidence set honestly. The new page is medium on five sources, not high, despite resting on two primary documents — a fresh page does not meet the independent-3-source bar. No page's confidence was raised this run.
• Cap respected. Eight gaps actioned against a 6–10 cap; 22 candidates found. Two searches and six fetches; no crawls; no failures.
• Lock held in file-only mode throughout, with heartbeats between phases. No git command was run from this session.

One-line summary

Cleared the 2026-09-07 report's own top-priority queue item and two more foundational sources, created Keith Ellison after three deferrals, and found that Kids Internet and Digital Safety Act (KIDS Act, H.R. 7757) recorded none of the multistate attorneys-general opposition to the bill — now folded in with its section citations; needing the curator's attention are a raw filename that asserts a signatory count the capture cannot support, a three-way contradiction in that count left deliberately unresolved, an unprocessed evening digest for the seventh consecutive run, an unauthenticated CourtListener connector for the sixth, and a confidence: high on single-source pattern now found in eight folders that this skill's daily cap cannot close.